metadata

license: apache-2.0
base_model: alignment-handbook/zephyr-7b-sft-full
tags:
  - alignment-handbook
  - generated_from_trainer
  - trl
  - dpo
  - generated_from_trainer
datasets:
  - HuggingFaceH4/ultrafeedback_binarized
model-index:
  - name: zephyr-7b-dpo-full
    results: []

zephyr-7b-dpo-full

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

Loss: 0.5028
Rewards/chosen: -0.9469
Rewards/rejected: -1.8932
Rewards/accuracies: 0.7656
Rewards/margins: 0.9463
Logps/rejected: -451.4661
Logps/chosen: -357.2325
Logits/rejected: 1.5731
Logits/chosen: 0.6530

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-07
train_batch_size: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 2
total_train_batch_size: 128
total_eval_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.1
num_epochs: 1

Training results

Training Loss	Epoch	Step	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.5545	0.21	100	-1.3212	-1.0287	-312.0799	-374.3159	0.5658	0.7188	-0.4953	0.6264	-1.1217
0.5026	0.42	200	-0.1773	0.5190	-352.4985	-439.3264	0.5202	0.7461	-0.8995	0.8723	-1.7718
0.5106	0.63	300	0.0862	0.9099	-342.0043	-424.9976	0.5104	0.7656	-0.7946	0.8339	-1.6285
0.4859	0.84	400	0.7818	1.7438	-360.3139	-457.9452	0.5031	0.7578	-0.9777	0.9803	-1.9580

Framework versions

Transformers 4.36.2
Pytorch 2.1.2+cu121
Datasets 2.14.6
Tokenizers 0.15.0